audio.cpp: C++/ggml Runtime for 12 Audio Models, 5x Faster TTS on CUDA

Core summary

One sentence to understand this update

audio.cpp, a new native C++/ggml inference framework, offers 12 audio models including Qwen3-TTS and PocketTTS, achieving up to 5x faster Text-to-Speech performance on CUDA compared to Python.

Impact & opportunity

What this could mean

Builders focused on audio AI can significantly enhance the performance and efficiency of their TTS and other audio applications by leveraging this optimized C++/ggml framework.